Every message an AI agent writes to a retail customer is a communication under the FCA Consumer Duty. It must meet the customer's information needs, be likely to be understood and support informed decisions. Agents write each message fresh, so firms cannot pre-approve them one by one. Approve the brief and sources instead, then test and sample what reaches customers.
What does the Duty require of communications?
The Consumer Duty's consumer understanding outcome asks firms to support customers' understanding. In substance, communications must:
- meet the information needs of the customers they are aimed at;
- be likely to be understood by those customers, given their characteristics, including any vulnerabilities;
- equip customers to make decisions that are effective, timely and properly informed;
- be tested, where appropriate, and monitored for whether they achieve this.
This applies to every channel, including one-to-one conversations. It sits alongside the long-standing principle that communications must be clear, fair and not misleading. Our explainer on what the Consumer Duty is covers the full framework.
Who does it apply to?
FCA-authorised firms whose products and services reach retail customers, across banking, lending, payments, insurance, investments and pensions. Our short answer to whether the Consumer Duty applies to AI agents is yes: an agent's message to a customer is the firm's message.
What changes when an agent writes the message?
Firms are used to approving communications before they go out: a letter template, a web page, an email campaign. Agents break that model in three ways.
Every message is unique. An agent composes a fresh reply each time, shaped by the customer's question and data. There is no single text to approve.
Messages are interactive. A customer can ask a follow-up, and the next answer must be consistent with the last. Understanding is built across a conversation, not a single page.
Agents can tailor, for better or worse. An agent can simplify language for a customer who is struggling, or switch to a shorter format on mobile. It can also bury a fee in the fourth paragraph, use jargon a customer did not ask for, or answer a cost question with a benefit.
So the control moves upstream and downstream. Upstream, approve what the agent may say: the brief, the sources of fact, the required elements of key messages. Downstream, test whether customers understand the result and sample what was actually sent.
What should an agent's communications always do?
A few rules make most of the difference:
- Put key information first. Costs, deadlines, consequences and the customer's options appear before supporting detail.
- Answer the question asked. If a customer asks what a missed payment will cost, the answer starts with the cost.
- Use the customer's words. Explain terms when they must be used; do not introduce jargon the customer did not.
- Be consistent. The agent's answer matches the firm's terms, website and letters. Facts come from approved sources, not from the model's memory.
- Time it right. Information arrives when the customer needs it to decide, not after.
- Notice confusion. Repeated questions, contradictory statements or signs of distress prompt a simpler explanation or an offer of a person.
The last point links to vulnerability. Customers whose capability is affected by health, life events or circumstances may need more. Our piece on vulnerable customers and AI agents covers how to spot that.
How do you meet the Duty for agent-written messages?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Meets information needs | Brief lists required elements for key topics (fees, deadlines, rights) | Approved brief versions |
| Accurate and consistent | Agent draws facts only from approved, current sources | Source list with version history |
| Likely to be understood | Readability and plain-language checks on sampled messages | Sample review results |
| Tested with customers | Comprehension testing of agent answers on key topics | Test results and changes made |
| Clear, fair and not misleading | Second-line review of sampled conversations against a checklist | Review records and corrections |
| Monitored over time | Repeat contacts, complaints and confusion signals tracked | Outcome monitoring reports |
| Changes controlled | Brief or source changes re-tested before release | Change records with test results |
Sampling is a form of surveillance, and it should be designed as one. Our explainer on communications surveillance describes how firms structure it. A compliance agent can help by pre-screening conversations for missing key information, but a person reviews the findings, in line with our guide to AI compliance agents.
What will the FCA or internal audit ask?
- How do you approve what agents say when every message is different?
- Show us an agent conversation about fees. Was the cost clear and early?
- How have you tested whether customers understand agent answers?
- How do you keep agent answers consistent with your terms and website?
- What do repeat contacts and complaints tell you about agent communications?
- What changed after your last review?
What does the communications outcome not cover?
It does not decide whether a product offers fair value or suits its target market; those are separate outcomes. It does not cover the routes customers use to get help, which falls under consumer support; see Consumer Duty and AI agents in customer service. It does not cover data protection notices, which follow UK GDPR. And it does not apply to non-retail customers.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Through Pryme Context Protocol, every fact an agent states arrives stamped with its source, so a sampled message can be checked against the approved terms it relied on. Brief and source changes are versioned, and Attest re-tests the agent on real conversations before a change goes live.
Questions people ask
Must every agent message be approved before it is sent?
No. That would defeat the purpose of an agent. Approve the brief, sources and required elements, test comprehension, and sample what is sent.
Can an agent simplify our terms for a customer?
It can explain them in plain words, as long as the explanation is accurate and does not leave out anything material. Test simplified explanations with customers.
Do agent-written emails count the same as chat?
Yes. Any communication to a retail customer is in scope, whatever the channel or format.